yaxunl created this revision.
yaxunl added reviewers: tra, ashi1.
Herald added a subscriber: jholewinski.
Add a new gpu arch gfx906+sram-ecc for HIP.
It is similar to gfx906 but implies target feature sram-ecc enabled,
whereas gfx906 implies target sram-ecc disabled. Corresponding
option -mattr=[+|-]sram-ecc will be passed to llc based on the
gpu arch.
https://reviews.llvm.org/D59863
Files:
include/clang/Basic/Cuda.h
lib/Basic/Cuda.cpp
lib/Basic/Targets/NVPTX.cpp
lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
lib/Driver/ToolChains/HIP.cpp
test/Driver/hip-toolchain-sram-ecc.hip
Index: test/Driver/hip-toolchain-sram-ecc.hip
===================================================================
--- /dev/null
+++ test/Driver/hip-toolchain-sram-ecc.hip
@@ -0,0 +1,38 @@
+// REQUIRES: clang-driver
+// REQUIRES: x86-registered-target
+// REQUIRES: amdgpu-registered-target
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN: -x hip --cuda-gpu-arch=gfx906 --cuda-gpu-arch=gfx906+sram-ecc \
+// RUN: %s 2>&1 | FileCheck %s
+
+// CHECK: [[CLANG:".*clang.*"]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
+// CHECK-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
+// CHECK-SAME: "-emit-llvm-bc"
+// CHECK-SAME: {{.*}} "-target-cpu" "gfx906"
+
+// CHECK: [[OPT:".*opt"]] {{".*-gfx906-linked.*bc"}} "-mtriple=amdgcn-amd-amdhsa"
+// CHECK-SAME: "-mcpu=gfx906"
+// CHECK-SAME: "-o" [[OPT_906_BC:".*-gfx906-optimized.*bc"]]
+
+// CHECK: [[LLC: ".*llc"]] [[OPT_906_BC]]
+// CHECK-SAME: "-mtriple=amdgcn-amd-amdhsa" "-filetype=obj"
+// CHECK-SAME: {{.*}} "-mcpu=gfx906"
+// CHECK-SAME: "-mattr=-sram-ecc" "-o" {{".*-gfx906-.*o"}}
+
+// CHECK: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
+// CHECK-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
+// CHECK-SAME: "-emit-llvm-bc"
+// CHECK-SAME: {{.*}} "-target-cpu" "gfx906"
+
+// CHECK: [[OPT]] {{".*-gfx906\+sram-ecc.*bc"}} "-mtriple=amdgcn-amd-amdhsa"
+// CHECK-SAME: "-mcpu=gfx906"
+// CHECK-SAME: "-o" [[OPT_906ECC_BC:".*-gfx906\+sram-ecc.*bc"]]
+
+// CHECK: [[LLC]] [[OPT_906ECC_BC]]
+// CHECK-SAME: "-mtriple=amdgcn-amd-amdhsa" "-filetype=obj"
+// CHECK-SAME: {{.*}} "-mcpu=gfx906"
+// CHECK-SAME: "-mattr=+sram-ecc" "-o" {{".*-gfx906\+sram-ecc.*o"}}
+
+// CHECK: {{".*clang-offload-bundler"}}
+// CHECK-SAME: "-targets=host-x86_64-unknown-linux,hip-amdgcn-amd-amdhsa-gfx906,hip-amdgcn-amd-amdhsa-gfx906+sram-ecc"
Index: lib/Driver/ToolChains/HIP.cpp
===================================================================
--- lib/Driver/ToolChains/HIP.cpp
+++ lib/Driver/ToolChains/HIP.cpp
@@ -31,6 +31,14 @@
namespace {
+void stripFeatureString(StringRef &Arch) {
+ Arch = Arch.take_until([](char c) { return c == '+'; });
+}
+
+StringRef getFeatureString(StringRef Arch) {
+ return Arch.drop_until([](char c) { return c == '+'; });
+}
+
static void addBCLib(Compilation &C, const ArgList &Args,
ArgStringList &CmdArgs, ArgStringList LibraryPaths,
StringRef BCName) {
@@ -114,6 +122,7 @@
Compilation &C, const JobAction &JA, const InputInfoList &Inputs,
const llvm::opt::ArgList &Args, llvm::StringRef SubArchName,
llvm::StringRef OutputFilePrefix, const char *InputFileName) const {
+ stripFeatureString(SubArchName);
// Construct opt command.
ArgStringList OptArgs;
// The input to opt is the output from llvm-link.
@@ -162,11 +171,24 @@
Compilation &C, const JobAction &JA, const InputInfoList &Inputs,
const llvm::opt::ArgList &Args, llvm::StringRef SubArchName,
llvm::StringRef OutputFilePrefix, const char *InputFileName) const {
+
+ // Only two gpu archs for gfx906 are accepted: gfx906 and gfx906+sram-ecc.
+ // gfx906 implies sram-ecc off, whereas gfx906+sram-ecc implies sram-ecc on.
+ StringRef Feature = getFeatureString(SubArchName);
+ stripFeatureString(SubArchName);
+ assert(Feature.empty() ||
+ (SubArchName == "gfx906" && Feature == "+sram-ecc"));
+ if (SubArchName == "gfx906" && Feature.empty())
+ Feature = "-sram-ecc";
+
// Construct llc command.
ArgStringList LlcArgs{InputFileName, "-mtriple=amdgcn-amd-amdhsa",
"-filetype=obj", "-mattr=-code-object-v3",
Args.MakeArgString("-mcpu=" + SubArchName)};
+ if (!Feature.empty())
+ LlcArgs.push_back(Args.MakeArgString("-mattr=" + Feature));
+
// Extract all the -m options
std::vector<llvm::StringRef> Features;
handleTargetFeaturesGroup(
@@ -372,6 +394,7 @@
}
if (!BoundArch.empty()) {
+ stripFeatureString(BoundArch);
DAL->eraseArg(options::OPT_march_EQ);
DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ), BoundArch);
}
Index: lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
===================================================================
--- lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -4951,6 +4951,7 @@
case CudaArch::GFX902:
case CudaArch::GFX904:
case CudaArch::GFX906:
+ case CudaArch::GFX906_SRAM_ECC:
case CudaArch::GFX909:
case CudaArch::UNKNOWN:
break;
@@ -5004,6 +5005,7 @@
case CudaArch::GFX902:
case CudaArch::GFX904:
case CudaArch::GFX906:
+ case CudaArch::GFX906_SRAM_ECC:
case CudaArch::GFX909:
case CudaArch::UNKNOWN:
break;
Index: lib/Basic/Targets/NVPTX.cpp
===================================================================
--- lib/Basic/Targets/NVPTX.cpp
+++ lib/Basic/Targets/NVPTX.cpp
@@ -189,6 +189,7 @@
case CudaArch::GFX902:
case CudaArch::GFX904:
case CudaArch::GFX906:
+ case CudaArch::GFX906_SRAM_ECC:
case CudaArch::GFX909:
case CudaArch::LAST:
break;
Index: lib/Basic/Cuda.cpp
===================================================================
--- lib/Basic/Cuda.cpp
+++ lib/Basic/Cuda.cpp
@@ -109,6 +109,8 @@
return "gfx904";
case CudaArch::GFX906: // TBA
return "gfx906";
+ case CudaArch::GFX906_SRAM_ECC: // TBA
+ return "gfx906+sram-ecc";
case CudaArch::GFX909: // TBA
return "gfx909";
}
@@ -147,6 +149,7 @@
.Case("gfx902", CudaArch::GFX902)
.Case("gfx904", CudaArch::GFX904)
.Case("gfx906", CudaArch::GFX906)
+ .Case("gfx906+sram-ecc", CudaArch::GFX906_SRAM_ECC)
.Case("gfx909", CudaArch::GFX909)
.Default(CudaArch::UNKNOWN);
}
@@ -259,6 +262,7 @@
case CudaArch::GFX902:
case CudaArch::GFX904:
case CudaArch::GFX906:
+ case CudaArch::GFX906_SRAM_ECC:
case CudaArch::GFX909:
return CudaVirtualArch::COMPUTE_AMDGCN;
}
@@ -306,6 +310,7 @@
case CudaArch::GFX902:
case CudaArch::GFX904:
case CudaArch::GFX906:
+ case CudaArch::GFX906_SRAM_ECC:
case CudaArch::GFX909:
return CudaVersion::CUDA_70;
}
Index: include/clang/Basic/Cuda.h
===================================================================
--- include/clang/Basic/Cuda.h
+++ include/clang/Basic/Cuda.h
@@ -64,6 +64,7 @@
GFX902,
GFX904,
GFX906,
+ GFX906_SRAM_ECC,
GFX909,
LAST,
};
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
